ChatGPT Images 2.5 主打多轮改图不跑偏
OpenAI 称图像工具周产超 30 亿张，Images 2.5 重点解决「编辑漂移」：延迟降最多 50%，改一处不动其余，多轮修改仍保持人物与品牌细节。
OPENAI'S IMAGE TOOLS NOW PRODUCE MORE THAN 3 BILLION IMAGES A WEEK That scale explains why ChatGPT Images 2.5 is focused less on a flashy one-shot demo and more on the expensive part of real creative work: revision. OpenAI says the new model cuts generation latency by up to 50%, preserves people and objects from reference photos more reliably, and changes only the requested element while keeping the rest of the composition intact. The technical enemy is edit drift. Ask an image model to change a jacket, then a background, then one line of text, and the face, layout or brand details can mutate along the way. Images 2.5 is designed to carry those constraints through longer multi-turn sessions instead of rebuilding the picture after every instruction. Sketches, image comments, templates and shareable prompts move the workflow closer to art direction than prompt gambling. The claims are OpenAI's until independently tested, but the product direction is clear: the winning image model may be the one that survives the fifth correction, not the one that makes the prettiest first draft. Lesson 21: Why image editing is a memory problem
这条是引用着另一条写的：x.com/corvusxbt/status/2089688741249945812